I’ve not myself experimented with 5ttgen on MP2RAGE data, so there may yet be some nuances that I’ve not yet come across.
5ttgen freesurfer isn’t very commonly used: while it is based on FreeSurfer segmentation, which is appealing, it’s a simple voxel-by-voxel mapping of binary tissue masks, so there are no partial volume estimates, that’s why it appears so jagged. Indeed it was intended primarily as an exercise in demonstrating that multiple algorithms can be proposed and implemented and distributed for a given image processing task…
Version 3.0.0 comes with 5ttgen hsvs, which is described here and is my preferred solution.